hvantk-resource-expression-atlas - SKILL.md Agent Skill

name: hvantk:resource-expression-atlas description: Build an AnnData object from an EBI Expression Atlas baseline bulk-RNA-seq experiment (TPM expression matrix + SDRF sample metadata). status: provisional backend: anndata domain: transcriptomics

EBI Expression Atlas resource skill

Read hvantk/skills/_conventions/SKILL.md first. This skill assumes every convention there.

1. Status & scope

Status: provisional. Builder, downloader, dataset class, and downloader tests are in place under the plugin folder; round-trip builder snapshots are NOT yet seeded (no fixture, no schema.json, no sample_rows.json) -- writing the first round-trip test is follow-up work tracked alongside this skill.
In scope: any single Expression Atlas baseline bulk-RNA-seq experiment with a gene-centric TPM matrix (genes x samples) and a paired SDRF metadata file, converted to an AnnData object keyed samples x genes.
Out of scope: scRNA-seq cell-level matrices (use the UCSC Cell Browser plugin); differential expression matrices; cross-accession multi-experiment merging; gene-symbol / accession normalization (downstream).

2. Source identity

Provider: EBI Expression Atlas (https://www.ebi.ac.uk/gxa).
Catalog entry: the per-accession dataset list lives in hvantk/skills/expression_atlas/catalog/datasets.json (filterable via data_source == "Expression_Atlas"). Browse with hvantk catalog list --data-source Expression_Atlas or hvantk catalog show E-GTEX-8. A top-level expression-atlas provider-level catalog entry remains a TODO; when added, point the plugin manifest's source.catalog_ref at it and remove this note.
Per-accession download URLs derive from EXPRESSION_ATLAS_BASE_URL in hvantk/skills/expression_atlas/shared/constants.py and the FTP prefix /pub/databases/microarray/data/atlas/experiments/<accession>/.

3. Backend choice + reasoning

backend: anndata, domain: transcriptomics. Per _conventions § 3 "Expression (dense, Hail-friendly) → MatrixTable" vs "Expression (sparse, single-cell) → anndata h5ad" — Expression Atlas baseline bulk-RNA-seq sits between the two: matrices are dense floats but typically small enough (tens of thousands of genes x hundreds of samples) that the scanpy / AnnData ecosystem is the natural fit. Downstream consumers (cross-tissue baseline expression queries, tissue-specificity scoring via tspex) operate on AnnData. Hail offers no advantage at this scale.

4. Raw format & gotchas

The exact raw-format gotchas live where the parser does:

Expression TSV parser + AnnData assembly: hvantk/skills/expression_atlas/shared/expression_atlas.py (create_anndata_from_expression_atlas).
SDRF long → wide reshape: same file (_import_sdrf, _reshape_sdrf_long_to_wide_format, convert_sdrf_to_dataframe).
High-level builder narrative: docstring on build_expression_atlas in hvantk/skills/expression_atlas/builder.py.

Stable notes:

Expression matrix is gene-centric (rows = genes, columns = samples). The first two columns are Gene ID and Gene Name; everything after that is one float column per sample. The builder transposes to AnnData's samples-as-obs convention.
SDRF is tab-separated, no header, and condensed-long: each row is (accession, unused, sample_id, column_type, column_name, column_value). The unused column is dropped; column_type is either characteristic or factor; duplicates on (sample_id, column_name) keep the last value (see _reshape_sdrf_long_to_wide_format).
Column names from SDRF are normalized: spaces → underscores, parentheses stripped (organism_part_(group) → organism_part_group). Downstream obs column names follow this rule.
TODO: enumerate per-accession quirks (e.g. mixed-type factor columns, missing SDRF rows, multi-pipeline TPM variants) once the round-trip fixture is seeded. Add specific cases here as they are encountered.

5. Output contract

Object: anndata.AnnData optionally saved to <output_path>.h5ad.
Shape: obs = samples, var = genes. X is float32 (genes-x-samples after transpose).
obs: indexed by sample_id; columns are SDRF characteristics / factors after the long → wide reshape (e.g. organism, tissue, cell_type, ...).
var: indexed by gene_id. Includes a Gene Name column when the source TSV had one.
uns["column_summary"]: per-obs-column summary annotated by annotate_column_summary_ad.
Provenance: stamped on the returned ExpressionMatrix via ctx.provenance(schema_id="expression-atlas-dataset-v1"); persisted by the platform as a sidecar .provenance.json.

6. hvantk integration points

Builder: build_expression_atlas (signature (parsed_input, ctx, **params) -> ExpressionMatrix) in hvantk/skills/expression_atlas/builder.py.
SDRF / matrix helpers: hvantk/skills/expression_atlas/shared/expression_atlas.py.
Dataset / collection classes: ExpressionAtlasDataset, ExpressionAtlasDatasetCollection in hvantk/skills/expression_atlas/shared/datasets.py.
Downloader CLI: download_experiments in hvantk/skills/expression_atlas/cli.py (registered as hvantk expression-atlas-download and also re-bound under hvantk download expression-atlas).
Lifecycle entry point: download_dataset in hvantk/skills/expression_atlas/cli.py.
Build CLI: hvantk reprocess expression-atlas:dataset --raw-dir <dir> --output <path>.h5ad (delegates to the plugin builder; pass builder kwargs via --plugin-arg key=value).
Plugin manifest: hvantk/skills/expression_atlas/plugin.yaml (drives loader registration; compound dataset key expression-atlas:dataset).
Tests: hvantk/skills/expression_atlas/tests/ (downloader unit + drift-probe sanity present; builder round-trip TODO).

Read the existing files at these paths as ground truth for shape. This skill does not restate code.

7. Workflow steps

When invoked to build or update a single Expression Atlas experiment:

Resolve raw paths. Either download via hvantk expression-atlas-download --accession <E-XXXX-N> --download_path /tmp/atlas, or via the recipe system: hvantk reprocess expression-atlas:dataset (lifecycle download → builder).
Build: hvantk reprocess expression-atlas:dataset --raw-dir <dir> --output <out>.h5ad.
- The builder parses the SDRF, transposes the expression matrix, attaches per-sample metadata into obs, annotates provenance, and writes .h5ad.
Validate: TODO — once the round-trip fixture is seeded, run pytest hvantk/skills/expression_atlas/tests. Until then, the offline downloader unit tests + drift-probe placeholder test are what guard this plugin.

8. Update playbook

TODO. This section will be fleshed out once per-accession drift detection lands (see § 2 catalog note and the drift-probe placeholder in hvantk/skills/expression_atlas/drift_probe.py). Expected shape:

For each tracked accession in hvantk/skills/expression_atlas/catalog/datasets.json (filter data_source == "Expression_Atlas"), re-run the per-accession HEAD probe; flag accessions whose Last-Modified or Content-Length changed.
Re-download flagged accessions, rebuild via hvantk reprocess expression-atlas:dataset, and diff the new AnnData against the snapshotted shape / obs columns.
If the SDRF column set changed, document the new factor in § 4.

9. Validation contract

Per _conventions § 9:

fixture: hvantk/skills/expression_atlas/tests/testdata/raw/expression-atlas/ (directory present, not yet seeded). The first round-trip test will populate this with a small TSV + condensed-SDRF pair.
schema_snapshot: hvantk/skills/expression_atlas/tests/snapshots/schema.json (TODO — created on first --regenerate-snapshots run).
row_snapshot: hvantk/skills/expression_atlas/tests/snapshots/sample_rows.json (TODO — same).
test_command: pytest hvantk/skills/expression_atlas/tests.

The plugin manifest already declares these paths so the loader contract holds. The downloader unit tests + drift-probe placeholder test pass today; the builder round-trip is the gap to close in a follow-up PR.

Snapshot status: schema.json and sample_rows.json have NOT yet been seeded for this plugin. On first round-trip run in a hail-enabled environment, use pytest hvantk/skills/expression_atlas/tests/test_builder.py --regenerate-snapshots to bootstrap them, then commit. Until seeded, the round-trip test cannot verify output against a fixed schema.